The law of AI knowledge distillation

Abstract

Knowledge distillation—the practice of training a compact student model on the outputs of a larger teacher model—has emerged as the fastest‐growing technique for replicating and disseminating frontier artificial intelligence (AI) capabilities. Because distillers can achieve near‐parity with incumbents at a fraction of the cost, the practice simultaneously promises dramatic gains in competition and efficiency, while threatening to upend the incentive structure that finances multibillion‐dollar AI research. Existing U.S. law offers only fragmentary guidance: copyright might not protect AI outputs at all; patent law offers limited, implementation‐specific protection; trade‐secret doctrine is pierced by the reverse‐engineering exception; contract law's reach is powerful but uneven; and antitrust principles loom in the background yet remain doctrinally underdeveloped. As a result, both developers and distillers confront a haze of uncertainty that chills beneficial innovation, invites strategic litigation, and leaves core normative questions unanswered. This Article provides the first comprehensive treatment of the law of AI knowledge distillation. It begins by mapping the terrain across copyright, patent, trade‐secret, contract, and antitrust doctrines, exposing the fault lines and unresolved conflicts that make distillation litigation unpredictable. It then draws on utilitarian, deontological, Lockean, and Rawlsian theories to evaluate whose interests the law ought to protect and why. Building on these positive and normative insights, the Article then proposes a balanced, modular framework that (1) classifies distillation into authorized, independent, and illicit forms; (2) legalizes good‐faith reverse‐engineering of publicly available model outputs; (3) imposes a time‐limited compulsory‐licensing regime on direct commercial cloning; (4) preserves broad exemptions for research, accountability, and transformative use; and (5) augments these rules with transparency, privacy, and safety safeguards. By articulating clear, principled boundaries around permissible and impermissible distillation, the Article charts a path that rewards foundational innovation without entrenching monopolies, facilitates robust competition and oversight, and ensures the more equitable distribution of the social surplus that AI generates.